Home Assistant 2026.10 landed on 7 October, and the headline isn’t the new vector maps or the tidier automation editor. It’s a small card on the settings page.

The old AI tasks page under Settings > System is now just called AI, and it has gained a Model Context Protocol card. In the release notes’ own words: “Turn the server on in a single click, and the card lists the URLs to paste into your AI app, each with a button to copy it.” (Home Assistant 2026.10 release notes)

That means the gap between “I have a Home Assistant box in my HDB” and “Claude or ChatGPT can switch off my aircon” is now about thirty seconds. The plumbing has been there since Home Assistant shipped its MCP Server integration. What 2026.10 removes is the friction, and friction was doing a fair bit of safety work.

Hooking a frontier AI model up to your flat isn’t a bad idea. It can be very useful. But the defaults decide what the AI can touch, and those defaults weren’t written with a Singapore flat in mind. Here’s what actually changes, what gets shared, and the ten minutes of setup I’d do before pasting that URL anywhere.

What MCP Actually Is (in One Paragraph)

The Model Context Protocol is an open standard that lets an AI app call “tools” on another system. Think of it as a USB port for AI assistants. Home Assistant’s MCP Server integration turns your instance into one of those systems. It offers the AI tools (to perform actions through Home Assistant), prompts (instructions on how to use those tools), and a read-only resource with a live snapshot of your home’s state.

The AI app runs on someone else’s servers. Your Home Assistant runs in your flat. MCP is the conversation between them.

What’s New in 2026.10, Specifically

Three things matter here:

  1. One-click server. Before this release you had to know the MCP Server integration existed, add it, and then work out which URL to give your AI app. Now it’s a toggle in Settings > System > AI, with copy buttons for the URLs.
  2. Auto-discovery the other way round. Apps that run their own MCP server are now discovered automatically by Home Assistant’s Model Context Protocol (client) integration, so connecting to them “only takes a confirmation.” This is the reverse direction: Home Assistant’s own AI using outside tools.
  3. Better music search. A small one. Ask your AI to play something and it now gets a list of results and picks the best match, instead of grabbing the first hit.

The same release also confirmed that Home Assistant Cloud is being renamed Home Assistant Link, with the switch planned for the December release. Nabu Casa says it has moved its remote-access servers to bare metal, added more servers worldwide, and shipped a latency checker that pings several servers when Home Assistant boots and picks the fastest. That matters more than it sounds, for reasons I’ll get to.

The Catch: Cloud AI Needs a Door Into Your Flat

This is the bit the one-click card doesn’t spell out. Claude, ChatGPT and other web-based AI apps don’t run on your network. The MCP Server docs are blunt about it: for cloud-based clients, “your Home Assistant instance must be accessible from the internet.”

For most Singapore households that leaves three options:

  • Home Assistant Cloud (soon Link). The easy route. It’s US$6.50 a month or US$65 a year on Nabu Casa’s pricing page, so roughly S$8–9 a month at current exchange rates. You get remote access without touching your router, plus encrypted backups, and the money funds Home Assistant’s development.
  • A tunnel you run yourself, like Cloudflare Tunnel. It’s free, but now you’re responsible for a public endpoint into your home. That’s fine if you know what you’re doing. If you’re not sure, you probably don’t.
  • Port forwarding on your ISP router. Please don’t. And on fibre plans that put you behind carrier-grade NAT, it won’t even work.

If you already pay for Home Assistant Cloud, the new latency checker is a quiet win. Remote access from Southeast Asia has always meant your traffic takes a round trip to wherever the nearest relay happens to be. Nabu Casa hasn’t published a server map with this release, so I can’t promise there’s a box in Singapore. But “pick the fastest server automatically” beats “you get whatever you’re assigned”, especially when the AI is making several tool calls to answer one question.

One more requirement: by default, only administrator accounts can use the /api/mcp endpoint. Web apps sign in with OAuth. Command-line tools like Cursor need a long-lived access token instead. Either way, whatever you connect is effectively logging in as an admin. Treat it that way.

What the AI Can See and Touch by Default

This is the part worth slowing down for.

The MCP server passes your AI the same “LLM APIs” Home Assistant uses for its own voice assistant, and the main one is the Assist API. Its rule is simple and mostly sensible: the AI can only control or report on entities that are exposed to Assist. You manage that list under Settings > Voice assistants > Expose.

The question is what’s on that list before you’ve touched it. I went through Home Assistant’s source to check. Out of the box, Assist auto-exposes entities from these domains:

  • climate (your aircon, via a Wi-Fi adapter or IR blaster)
  • cover (curtains, blinds, and also gate and garage-door motors)
  • fan, humidifier, light, media_player, scene, switch, todo, vacuum and water_heater

It also exposes door, window, opening, lock-status, motion and presence binary sensors, plus temperature, humidity, CO, CO2, PM2.5, PM10, AQI and VOC sensors. And the setting that automatically exposes new entities is on by default for Assist.

There’s good news here too. The lock domain is not on that list. Your front-door smart lock isn’t controllable by Assist, or by an AI going through it, unless you expose it yourself. Neither are alarm panels or cameras. Entities flagged as config or diagnostic, or hidden by the system, are also left out.

Now translate that default into a typical Singapore flat.

The things that are probably fine

Lights, aircon, the robot vacuum, curtains, the living-room speaker, scenes. If the AI misreads you and switches the bedroom aircon to 24°C, nobody gets hurt. This is exactly what an AI assistant is good for: “it’s getting stuffy, sort out the living room” works better with a model that can reason than with a rigid voice command.

The things I’d look at twice

  • Switches. “switch” covers everything from a lamp plug to the 20A switch on your storage water heater. A lot of HDB owners automate the heater with something like the Aqara H1 Smart 20A Switch so it only runs before showers. Having an AI switch it on by mistake won’t cause a disaster, but it will waste power at today’s tariffs. And if you’ve ever put anything on a smart plug that shouldn’t run unattended (a portable heater, a kettle, an iron), that plug is exposed to the AI by default.
  • Covers. Curtain motors are harmless. A motorised gate on a landed property is not. Home Assistant files both under the same domain.
  • Presence and motion sensors. These aren’t actions. They’re information, and that cuts the other way (see the next section).

The thing that should stay off

Your door lock. Home Assistant already leaves it off, so the only way it ends up exposed is if you turn it on. If you’ve added a smart lock like the Aqara U300 to Home Assistant so you can lock up remotely, that’s great. Exposing the lock entity would hand the AI both the status and the unlock button. If you just want it to answer “did I leave the door open?”, a door contact sensor, which is exposed by default, already tells it that. A model following instructions hidden in a web page or an email it’s summarising is a real class of attack, and “unlock the front door” is the one tool call you never want it to make by mistake.

What Leaves Your Flat

Home Assistant’s whole pitch is local control, and it’s a big reason people in Singapore pick it over cloud-first ecosystems. MCP doesn’t break that. Your automations still run locally, and your Zigbee, Thread and Matter devices still talk to your hub, not to OpenAI or Anthropic. If your fibre goes down, your lights still work.

What does leave is whatever the AI reads while it’s helping you. The MCP server can hand the model a live snapshot of every exposed entity. On default settings, that includes which rooms have motion, whether the main door is open, and what the presence sensor in the bedroom says. Put together, that’s a fairly precise record of when your flat is occupied, sent to a model provider’s servers every time you ask a question.

For most people that’s an acceptable trade, as long as it’s a choice you made on purpose. Two Singapore-specific reasons to be more careful:

  • If you employ a live-in helper, the motion and presence data in their part of the flat describes their movements as much as yours. We covered what the rules actually say about cameras and monitoring. Even where it’s legal, routing that data through a third-party AI is worth a second thought.
  • If you share the flat with family, remember the AI is acting through an admin account. Everyone’s presence data is in scope, not just yours.

If none of that sits right with you, the alternative is a local model. Home Assistant can talk to models running on your own hardware, and we’ve written up how to build a private, local voice assistant. It’s slower and less clever than a frontier model, but nothing leaves the flat. Just know that running a decent model locally needs more than a Home Assistant Green, and hardware prices haven’t been kind this year.

The Ten-Minute Setup I’d Actually Do

Before you hit that toggle:

  1. Audit the Expose tab. Go to Settings > Voice assistants > Expose, and walk through every switch and cover on the list. Unexpose anything where a wrong action costs money or affects safety: water heater, high-load plugs, gates.
  2. Turn off auto-exposure for new entities. Otherwise the next smart plug you pair goes straight onto the AI’s list without you ever deciding it should.
  3. Leave locks, alarms and cameras alone. They’re unexposed by default. Keep it that way.
  4. Narrow the APIs. The MCP Server integration has an “Expose all LLM APIs” option that’s on by default. It hands over every LLM API you have installed, including future ones that custom integrations might add. Turn it off and tick only Assist unless you know you need something else.
  5. Give each entity a clear name and aliases. “Master bedroom aircon” works far better than “climate.living_room_2” when an AI is guessing what you mean. HDB layouts with a “common bedroom 2” and a “study that’s actually a store room” need this more than most.
  6. Sort out remote access properly. Use Home Assistant Cloud or a tunnel you understand. If your IoT gear is on its own network segment, this is a good time to check the segmentation still holds.
  7. Paste the URL into your AI app and test with read-only questions first. “Which windows are open?” and “What’s the humidity in the bathroom?” are good first prompts. Only once those come back sensible should you try “turn off everything in the living room.”

Custom connector support and the plans that include it change often on both Claude and ChatGPT, so check your app’s current connector settings rather than trusting a screenshot from last month.

Is It Actually Useful?

Yes, and mostly not for the reason people expect.

Switching lights on and off by chat is a gimmick. The useful part is having an AI that can see your home while you ask it questions. “Why did the bathroom fan run for three hours last night?” “Which room was warmest this afternoon?” “Write me an automation that closes the curtains when the living room hits 30°C, but only on weekdays.” Home Assistant has always been powerful, and also hard to configure. An AI that can read your actual entity names and states, instead of giving generic YAML advice, closes a lot of that gap.

If you’ve just bought your first Home Assistant box, this is also a decent way to learn your own setup. Ask the AI to list what’s exposed and you’ll often find a dozen entities you forgot existed. Our Matter network map walkthrough from last month’s release pairs well with it.

The Verdict

Home Assistant 2026.10 makes connecting a cloud AI to your home as easy as it should be. The safety model underneath, an explicit list of exposed entities with locks left out by default, is a good one. But the defaults lean towards convenience. Switches, covers and water heaters are in, auto-exposure for new devices is on, and “Expose all LLM APIs” is ticked.

Spend ten minutes on the Expose tab before you spend thirty seconds on the toggle. Then go ask your AI why the bathroom is still at 85% humidity. In Singapore, the answer is usually “because it’s Singapore”, but at least now it can show you the graph.

If you’d rather someone else set up the hub, the exposure lists and the remote access, HomeSmart’s guide to choosing an Aqara hub is a good place to start on the hardware side. Aqara’s Matter-capable hubs bridge neatly into Home Assistant.